Manish Ghoshal · AI Engineer

Master of Data Science at Melbourne. I build AI systems across language, time-series, and cloud infrastructure. The parts that have to keep working after the demo ends.

PythonJavaC#SQLBashGitPyTorchTensorFlowTransformersPEFT / LoRAONNX RuntimeTensorRT / TritonFlashAttention-2Docker / KubernetesMLflowKubeflowZenMLDVCWeights & BiasesAWS SageMakerGCP Vertex AILangChainLangGraphCrewAIAutogenRAGPineconeWeaviateUnslothLLaVAMoESparkDelta LakeBigQueryPostgreSQLRedisFastAPIgRPCdbtTerraformGitHub ActionsKafkaSnowflakeSHAPLIMECaptumEvidentlyPrometheusGrafanaPythonJavaC#SQLBashGitPyTorchTensorFlowTransformersPEFT / LoRAONNX RuntimeTensorRT / TritonFlashAttention-2Docker / KubernetesMLflowKubeflowZenMLDVCWeights & BiasesAWS SageMakerGCP Vertex AILangChainLangGraphCrewAIAutogenRAGPineconeWeaviateUnslothLLaVAMoESparkDelta LakeBigQueryPostgreSQLRedisFastAPIgRPCdbtTerraformGitHub ActionsKafkaSnowflakeSHAPLIMECaptumEvidentlyPrometheusGrafana

about me

I'm drawn to the parts of AI that rarely make the demo: evals, data contracts, scaling, alerts, and the quiet fixes that decide whether a model survives contact with real users.

HEALTHCARE AI
QUANT FINANCE
RECSYS
NLP
COMPUTER VISION
FORECASTING

history

01 ▌Nov 2025 – Mar 2026
Research Computing / Data Engineering Intern
Walter and Eliza Hall Institute of Medical Research (WEHI)
PythonPathlibCLI
02 ▌Dec 2023 – Aug 2024
AI Engineer
Wolf Tech
PyTorchTime-SeriesDocker
03 ▌Jan 2024 – May 2024
AI Engineer Intern
Futurense Technologies
LangChainQLoRARAG
04 ▌Sep 2023 – Oct 2023
Mentee, Amazon ML Summer School
Amazon
Machine LearningAmazonSelective Program
05 ▌Jan 2023 – Apr 2023
Software Development Intern
Central Automation & IT, AM/NS India
C#ASP.NETOracle SQL

currently

Currently building toward production-grade ML: cleaner pipelines, sharper evals, better interfaces, and measurable behavior under real constraints.

selected work

proc · live

WardSignal

2026
ED Surge & Bed-Pressure Early-Warning Copilot

Early-warning copilot for hospital operations: forecasts emergency-department demand and bed-access pressure 4 hours, one shift, and 24 hours ahead, then turns each forecast into a calibrated Green/Watch/Action/Critical band with a recommended operational play.

lightgbmfastapireact
● live demo ↗
proc · live

Nebula

2025
Gas-Sensor Drift & Concentration Analysis Dashboard

Interactive web dashboard to explore gas-sensor signatures under sensor drift (ageing) and concentration (dose) changes, with controls to simulate scenarios.

reactvitemachine-learning
● live demo ↗
proc · idle

Cloud-Native Sentiment Analytics Pipeline

2025
Kubernetes · RabbitMQ · KEDA

Architected a Kubernetes ingestion + analytics pipeline (scrapers → RabbitMQ → Elasticsearch) processing 25k+ posts/day with near-real-time querying.

kubernetesrabbitmqkedaelasticsearch
proc · idle

Driver Drowsiness Detection

2024
CNN-LSTM · GAN

Real-time drowsiness detection using a CNN-LSTM architecture; achieved 90%+ accuracy through iterative training and evaluation.

cnn-lstmgancomputer-vision
proc · idle

MetaMetric

2023
AI-Driven Product Length Prognostication

Predicted product dimensions from noisy e-commerce catalogue text for the Amazon ML Challenge 2023 — a top-15 finish out of 1500+ teams.

nlplightgbmtransformers
proc · live

AgriLens.AI

2022
Plant Disease Diagnosis Demo

Full-stack plant-disease diagnosis demo (FastAPI backend + React frontend) with an image-upload workflow and clean success/error/retry states.

fastapireacttailwindcss
● live demo ↗

how i work

Research, then reuse.

Read the paper, find the repo, port what already works. Net-new code is the last resort.

Ship the unglamorous parts.

Eval harnesses, autoscaling, the cloud plumbing. The model is the easy part.

Keep it alive at 2am.

Dashboards, alerts, the pager. Research nobody can run is just a slide deck.

a working theory

The best model is the one that ships.

There's a person behind the deploys.

Things I do when I am not babysitting a training run. Shine a light on any of them.

In service

7 CupsTrained active-listening volunteer for emotional-health support.
Art of LivingSeva and hospitality across community programmes.
KathaWrote for underserved communities across India.

Under pressure

NCC officerSurvival, discipline, and leadership in the field.
State-level basketballRepresented my school; three CBSE cluster runs.
Debate championFirst place two years running, then script-writing too.

On the mic

Stanford Code in PlaceSection leader and learner, 2021 and 2023.
YouTube ↗A data-science channel, explaining the hard parts simply.
CheggComputer-science subject-matter expert for students.

boot sequence

manish : bash
The ManishOS desktop: a window-managed OS with terminal, MGBot assistant, games and easter eggs

the fun part

Don't read my portfolio.Boot it.

Everything above, but as a custom operating system: real terminal, a chat assistant, 3D, and a pile of easter eggs.

Boot ManishOS →